Identification and validation of prognostic markers for cuproptosis-related macrophage polarization genes in hepatocellular carcinoma

Abstract

In this study, we used single-cell sequencing data analysis to explore differentially expressed genes in the polarization process of macrophages in hepatocellular carcinoma. We then integrated these genes with cuproptosis-related genes (CRGs) to identify potential biomarkers. Through a rigorous screening process, including univariate Cox regression analysis and machine learning algorithms, we identified six key risk genes: GLIPR2, ANP32E, LIPT1, ALAD, ARSK, and PGAM1. These genes form the foundation of our prognostic risk prediction model. ROC curve analysis showed that these models had high specificity and accuracy in predicting prognostic characteristics, and Kaplan-Meier curve analysis showed that the survival rate of the low-risk group was significantly higher than that of the high-risk group. In addition, patients stratified by our model showed differences in tumor microenvironment, sensitivity to immunotherapy, and response to chemotherapy. After incorporating patient clinical data, we constructed a nomogram that further improved the accuracy of predicting patient survival. We further analyzed the expression characteristics and spatial distribution of these six risk genes in hepatocellular carcinoma through bulk transcriptomics, single-cell, and spatial transcriptomics data, and validated the expression of risk genes using qPCR. The construction of predictive models in this study helps clinicians to predict the overall survival of patients with hepatocellular carcinoma, which enables patient stratification and has the potential to help personalize patient treatment. The discovery of candidate tumor markers helps to identify potential targeted therapeutic options, which will play a key role in the diagnosis and treatment of hepatocellular carcinoma in the future.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-04373-3.

Introduction

In 2022 liver cancer was the sixth most common cancer and the third leading cause of cancer death worldwide, with 865,000 new cases and 757,940 deaths, and it ranked second in male cancer mortality at two to three times the rates in women [1]. In the majority of liver cancer cases, more than 90% are hepatocellular carcinoma (HCC), for which chemotherapy and immunotherapy are considered the best treatment options [2]. Although, the treatment of hepatocellular carcinoma has made significant progress in recent years, many challenges remain in the management of patients with advanced HCC [3]. In contrast, hepatocellular carcinoma continues to have a high mortality rate, mainly due to late diagnosis resulting from early detection failures [4]. Therefore, it is necessary to find relevant tumor markers to construct prognostic prediction models for patients, which is important for both clinical treatment and physician decision-making.

Cuproptosis, Tsvetkov et al. identified a new form of cell death, a mechanism of death that occurs during copper toxicity [5]. Aggregation of proteins involved in the lipoylated tricarboxylic acid (TCA) cycle during cuproptosis leads to destabilization of Fe-S cluster proteins and increased proteotoxic stress [6]. In contrast, accumulation of lipoylated proteins and instability of Fe-S cluster proteins trigger the onset of cell death [5]. Importantly, copper toxicity can be induced or inhibited by regulating intracellular copper ion concentration, a mechanism that has potential applications in cancer therapy [6]. Future studies of copper toxicity-induced therapy should focus on patient groups that may benefit from it, such as cancers with high mitochondrial metabolism levels, cancers with tumor stem cell features, and drug-resistant cancers, with the aim of providing new therapeutic strategies [7].

Both cuproptosis and macrophage polarization are crucial biological processes in tumors, and the interactions that exist between them deserve further scrutiny. In this study, we analyzed the CRGs involved in macrophage M1/M2 polarization in hepatocellular carcinoma using single-cell data analysis combined with extensive transcriptomic data, and constructed a model to predict patient prognosis. Distinguished from previous bulk-only signatures, our model was exhaustively validated across bulk–single-cell–spatial transcriptome dimensions, securing high accuracy in survival prediction. Beyond stratifying patients by risk, we revealed pronounced disparities in immune infiltration and in immunotherapy/chemotherapy sensitivity between high- and low-risk groups. In-depth analysis of six risk genes encompassed expression profiles across the whole transcriptome, single-cell, and spatial transcriptomics levels, including copy number variation analysis. This approach aims to uncover mutation patterns and guide drug prediction, supporting precision medicine and personalized treatment strategies.

Methods

Data source

The data for this study were obtained from the TCGA, GEO, ICGC, and GSA databases. The expression data and patient prognosis data for the model training set were derived from TCGA- LIHC (n = 464), while the validation set data were from ICGC LIRI-JP (n = 260). We performed relevant pathway scoring and differential gene analysis using a single-cell dataset of hepatocellular carcinoma from the GEO database (GSE149614, n = 18). Publicly available spatial transcriptome (ST) data are available in the Genome Sequence Archive (GSA) under accession number HRA000437, and more details about ST data are available at https://ngdc.cncb.ac.cn/gsa-human/browse/HRA000437.

Data analysis

Single-cell or Spatial transcriptome data processing and analysis

The initial gene expression matrix underwent preprocessing with the “Seurat” R package (version 4.4.0) to ensure data quality and prepare it for downstream analysis. The filtering criteria included retaining cells that expressed more than 300 genes, had at least 3 non-zero UMI counts across cells, contained less than 50% mitochondrial UMIs, exhibited more than 3% ribosomal protein content, and showed less than 0.1% hemoglobin content. Genes were kept in the dataset if they were detected in a minimum of 3 cells. Highly variable genes were selected as key features for principal component analysis (PCA), from which the top 30 significant components were used for Uniform Manifold Approximation and Projection (UMAP) to visualize the high-dimensional gene expression patterns in a two-dimensional space. To address batch effects, we applied the Harmony algorithm, ensuring that technical variability did not confound biological differences. Differentially expressed genes within each cell subpopulation were identified using the FindAllMarkers function, applying the Wilcoxon rank-sum test and adjusting p-values for multiple hypothesis testing using False Discovery Rate (FDR) correction. Cell clustering was performed with the FindClusters function, setting a resolution parameter of 0.5 to define distinct cellular populations. This clustering was further refined by aligning single-cell RNA sequencing data with reference datasets through the “SingleR” package, thereby assigning cell types based on expression profiles. Cell cycle heterogeneity and pathway activity scores were assessed using the AUCell algorithm, providing insights into the dynamic states and functional activities of individual cells. A differential expression analysis revealed 260 macrophage-specific DEGs, which were integrated with 19 literature CRGs and 170 MDRGs from MSigDB. Correlation network analysis, refined by gene nomenclature features to identify functionally related modules, yielded 36 cuproptosis-related macrophage polarization genes (CRMGs). The AddModuleScore function used to score cuproptosis and macrophage polarization pathways, and the lm function of the stats package were used to construct two regression models for evaluating the interaction between the two scores. Spatial transcriptome data were analyzed using the “spacexr” software package to associate cellular subpopulation annotations with spatial locations [8]. In order to study the spatial distribution of cuproptosis and macrophage polarization, we used the GSVA algorithm to score the two biological processes [9]. Based on the scores for cuproptosis and macrophage polarization pathways, the tissue was divided into areas with high cuproptosis scores, areas with high macrophage polarization scores, areas with high scores for both, and areas with low scores for both, using the median as the threshold.

Construction and evaluation of predictive models

Starting with an initial set of 36 CRMGs, we refined this list to 16 genes through univariate Cox regression analysis, which assesses the impact of each gene on patient survival independently. Subsequently, by utilizing the Lasso machine learning algorithm, which is known for its variable selection and regularization capabilities, we identified a more concise set of 6 risk genes that are most predictive of patient outcomes. The risk score for each patient was calculated using a linear combination of gene expression levels weighted by their corresponding coefficients from the Lasso model, following the formula: Riskscore = ∑(Expi*coefi), where Expi represents the expression level of the ith gene, and coefi is its associated coefficient. Patient risk groupings were based on median risk scores. To evaluate the differences in survival rates between high-risk and low-risk patient groups defined by these risk scores, we performed survival analysis using the Kaplan-Meier estimator, complemented by the log-rank test to statistically compare the resulting survival curves. For these analyses, as well as for constructing univariate and multivariate Cox regression models and visualizing results with forest plots, we utilized the “survival” and “survminer” R packages. In order to thoroughly assess the prognostic model’s performance and the nomogram’s accuracy, we employed several R packages including “rms” for regression modeling strategies, “pROC” and “timeROC” for ROC curve analysis to measure discrimination over time, and “dcurves” for clinical decision curve analysis (DCA) to evaluate net benefits across different probability thresholds. These evaluations aimed to gauge the prediction accuracy and clinical applicability of the Lasso regression model, the nomogram, and traditional clinical data. The detailed list of CRMGs considered in our study can be found in Table S1. The 260 differentially expressed genes are presented in Table S2.

Application of prognostic modeling

Patients were categorized into high-risk and low-risk groups according to their calculated risk scores. To analyze the differences in gene expression between these two groups, we utilized the “limma” package, a powerful tool for linear modeling of microarray data that is also widely applied to RNA sequencing data. Survival analysis was performed using the Kaplan-Meier estimator to compare survival outcomes between the high-risk and low-risk patient groups across hepatocellular carcinoma and other types of tumors following risk stratification. This method provides a visual representation of survival probabilities over time and allows for statistical comparison through tests like the log-rank test. For assessing variations in immune and stromal cell infiltration, we used the “CIBERSORT”, “MCPcounter”, and “quanTIseq” computational tools [10–12]. These packages estimate the abundance of various immune cell types from bulk gene expression data, each employing distinct algorithms and reference datasets to provide insights into the tumor microenvironment. The Tumor Immune Dysfunction and Exclusion (TIDE) database was consulted to forecast the likelihood of response to immunotherapies within each risk group, offering valuable information on potential therapeutic efficacy based on the tumor’s immune landscape [13]. Furthermore, to investigate the correlation between risk group classification and chemotherapeutic agent sensitivity, we applied the “oncoPredict” software package [14]. This tool integrates gene expression data with drug sensitivity information to predict how differentially expressed genes might influence treatment outcomes, thereby providing a basis for personalized therapy approaches.

Expression and mutation patterns of risk genes

To investigate the mutational landscape of genes associated with hepatocellular carcinoma, somatic mutation analysis was carried out using the “maftools” software package, which enabled us to examine and graphically represent the mutation patterns of these risk genes [15]. We further explored the expression profiles of a panel of 6 risk genes across normal tissues, lung cancer tissues, and diverse cellular subpopulations by leveraging bulk transcriptomics, single-cell RNA sequencing, and spatial transcriptomics datasets. This multi-faceted approach allowed for a comprehensive assessment of the functional roles and prognostic implications of these genes. For evaluating the prognostic significance of the risk genes in HCC, we utilized the GEPIA2 database, which offers a user-friendly interface for gene expression profiling and survival analysis [16]. To delve into the stage-specific expression changes of the ARSK gene within HCC patients, we turned to the UALCAN portal [17]. Moreover, immunohistochemical staining images from the Human Protein Atlas (HPA) database provided visual evidence of the expression levels of these risk genes in both normal lung tissue and lung cancer samples. In pursuit of potential therapeutic strategies, the Drug Gene Interaction Database (DGIdb) was employed to predict drug interactions for several of the identified risk genes, thereby highlighting possible targets for the treatment of HCC [18].

Validation using quantitative real-time quantitative PCR (qPCR)

To verify the expression of several risk genes, qPCR analysis was performed on the normal hepatocyte cell line HL-7702 (AbCells, Cat. No. AC254) and the hepatocellular carcinoma cell line HEPG2 (AbCells, Cat. No. AC111). Total RNA was extracted from all cell lines using the RNAprep Pure Cell/Bacteria Kit (TIANGEN, Beijing, China; Cat. No. DP430). Reverse transcription was performed using the PrimeScript RT Reagent Kit (Takara, Beijing, China; Cat. No. RR037A). qPCR was performed using the The qPCR was performed using SYBR Green qPCR mix (Beyotime, Beijing, China; Cat. No. D7260), and GAPDH was used as the internal reference for gene detection. qPCR conditions included pre-denaturation at 95 ℃ for 2 min, followed by denaturation at 95 ℃ for 15 s, annealing at 60 ℃ for 30 s, and elongation at 60 ℃ for 30 s, with a total of 40 cycles. The primer sequences are shown in Table S3. Three biological replicates were used, and the expression levels of the biomarkers in HEPG2 and HL-7702 cell lines were analysed by t-test, and a P value of less than 0.05 was considered as statistically significant.

Statistical analysis

The statistical processing, data handling, and graphical representation were conducted with R software, specifically version 4.2.3. Patient survival within different risk categories was assessed through the Kaplan-Meier method, while disparities between survival distributions were tested using the log-rank test. In comparing continuous variables across low-risk and high-risk cohorts, we applied either the Wilcoxon rank-sum test or the independent samples t-test, depending on whether the data followed a normal distribution. Analysis of categorical variables utilized the chi-squared test or Fisher’s exact test, whichever was more suitable. To account for multiple hypothesis testing, we adjusted p-values by applying the False Discovery Rate (FDR) correction. Relationships between variables were explored using Spearman’s rho for rank correlation. A threshold of P < 0.05 was set to determine statistical significance. Furthermore, the levels of significance in the findings are denoted as follows: *** for P < 0.001, ** for P < 0.01, and * for P < 0.05.

Result

A flowchart detailing the procedures of this study is shown in Fig. 1.

Fig. 1.

Flow chart of study design

Screening for genes associated with cuproptosis- related macrophage polarization

We obtained the hepatocellular carcinoma dataset GSE149614 from the GEO database. Following quality control and normalization, the dataset contained 61,486 cells and 24,017 genes. After automatic annotation of relevant cell subpopulations, we identified a total of 10 cell subpopulations, including adipocytes, B cells, CD8 + T cells, endothelial cells, fibroblasts, hepatocytes, macrophages, monocytes, NK cells and T cells (Fig. 2A). By evaluating the M1/M2 polarization pathways of macrophage subpopulations and dividing them into highly polarized and low-polarized groups based on median scores, a total of 260 differentially expressed genes (DEGs) were identified (Fig. 2B, Table S2). To further investigate macrophage polarization, we sourced gene sets from the MSigDB dataset. Given the limited sample size of the single-cell data, we performed correlation analysis using the TCGA-LIHC dataset among CRGs, the identified DEGs, and macrophage polarization genes. This identified 36 cuproptosis-related macrophage polarization genes (CRMGs) for subsequent univariate Cox regression analysis (Table S4). Univariate Cox regression analysis of these 36 genes identified 16 with significant prognostic value (P < 0.05) (Table S5). Subsequent refinement using the LASSO regression machine learning algorithm pinpointed six key genes as prognostic risk genes for hepatocellular carcinoma: GLIPR2, ANP32E, LIPT1, ALAD, ARSK, and PGAM1 (Fig. 2C,D, Table S6).

Fig. 2.

Screening of macrophage polarization genes associated with cuproptosis. A The UMAP visualization illustrates the clustering of HCC single-cell data according to cell type, where distinct hues correspond to individual cell types. B Within this diagram, each datapoint signifies a unique cellular subpopulation. The spectrum from blue to yellow reflects the gradient of AUCell scores for macrophage polarization, with blue indicating lower scores and yellow denoting higher ones. C Lasso coefficient paths track how each gene’s prognostic weight evolves with regularization; persistent, large coefficients highlight candidate drivers of patient outcome. D Ten-fold cross-validation curve pinpoints the penalty (λ) that yields the most parsimonious gene set while minimizing prediction error, defining the final signature

Risk modeling and evaluation

To evaluate the prognostic value of the six risk genes screened above, we constructed a multigene combination risk score model based on their expression. The risk score was calculated as: Riskscore = (0.0326 * expression of GLIPR2) + (0.1586 * expression of ANP32E) + (0.0740 * expression of LIPT1) + (− 0.2013 * expression of ALAD) + (0.2125 * expression of ARSK) + (0.0141 * expression of PGAM1). Based on the median risk score, patients from the TCGA cohort were stratified into high- and low-risk groups. Kaplan-Meier survival analysis demonstrated a significant survival difference between these two groups in the TCGA dataset (P < 0.05) (Fig. 3A). This significant survival difference was also validated in the ICGC dataset (P < 0.05) (Fig. 3B). ROC curve analysis revealed that the risk scores provided higher predictive sensitivity compared to clinical factors alone (Fig. 3C). Risk factor distribution maps illustrated patient distribution across risk categories (Fig. 3D). Furthermore, multifactor Cox regression analysis indicated that risk scores are independent predictors of survival, surpassing other clinical information (Fig. 3E).

Fig. 3.

Evaluation of the risk model. A The survival analysis of the training set reveals how survival probabilities evolve over time for patients categorized into high-risk and low-risk groups. A statistically significant difference in survival probabilities exists between these two groups (P < 0.001). B The validation set’s survival analysis indicates a notable divergence in survival probabilities between the high-risk and low-risk patient groups over time (P < 0.05), reinforcing the initial findings. C To evaluate the accuracy of the survival prediction model, ROC curves were constructed, yielding AUC values of 0.680 at one year, 0.676 at three years, and 0.687 at five years post-diagnosis. D Risk-based model for prognostic assessment. High-risk patients showed significant association with poorer overall survival (P < 0.001). Key risk-associated genes exhibited differential expression between high- and low-risk groups, validating the model’s biological basis. E Multivariate analysis further confirmed that the risk score is a strong prognostic predictor independent of clinical characteristics such as age, gender, and TNM staging (P < 0.001)

Risk modeling applications

To further characterize the biological differences between the high- and low-risk groups, we systematically profiled their tumor immune microenvironment (TIME). We assessed immune cell infiltration in both groups using the CIBERSORT and MCPcounter algorithms. MCPcounter analysis revealed significantly higher infiltration of fibroblasts in the high-risk group. Conversely, the high-risk group exhibited lower levels of CD8 + T cells, cytotoxic lymphocytes, NK cells, neutrophils, and endothelial cells compared to the low-risk group (Fig. 4A). CIBERSORT results indicated elevated abundances of T follicular helper cells, T regulatory cells (Tregs), and macrophages in the high-risk group. In contrast, resting NK cells, monocytes, and resting mast cells were less abundant in the high-risk group (Fig. 4B). GSVA enrichment analysis was performed on the two groups, and significant differences in the pathways enriched in each group were found, such as fatty acid metabolism, cholesterol homeostasis, and WNT β-catenin signaling pathway (Fig. 4C). We analyzed immune checkpoint genes in tumor patients from previous studies and identified significant expression differences in 57 genes (Fig. 4D, Table S7). Drug sensitivity analysis predicted that the high- and low-risk groups exhibited distinct responses to chemotherapeutic agents, including oxaliplatin and sorafenib (Fig. 4E, Table S8). The extent of M1 Macrophages infiltration in the two risk groups was calculated using the quanTIseq algorithm and box plots were used to show differences between groups (Fig. 4F). Analysis of patients’ immune responsiveness revealed that high-risk patients had poor immune responsiveness (Fig. 4G). There were significant differences in survival prognosis, immune cell infiltration, and immunotherapy responsiveness between patients in the high- and low-risk groups in the validation set (Figure S1).

Fig. 4.

Utilizing Risk Modeling to Assess Patient Immune Infiltration, Predict Immunotherapy Response, and Evaluate Chemotherapy Efficacy. A MCPcounter was utilized to assess the variations in immune cell infiltration between patients categorized into high-risk and low-risk groups. B CIBERSORT was employed to examine disparities in both immune and stromal cell infiltration between the high-risk and low-risk patient cohorts. C GVSA analysis was conducted to explore differences in pathway enrichment between the high-risk and low-risk groups of patients. D Expression levels of several immune checkpoint genes were compared in the high- and low-risk groups. E OncoPredict analysis was performed to evaluate the differential sensitivity to chemotherapeutic agents exhibited by the two risk-stratified patient groups. F The quanTIseq algorithm was applied to compare M1 tumor-associated macrophage infiltration levels in patients from the high-risk and low-risk categories. G A violin plot illustrates the distribution of TIDE scores across high-risk and low-risk patient groups, indicating variations in immunotherapy response efficacy between these two groups

Nomogram construction and evaluation

To improve prognostic prediction, we constructed a nomogram that integrated the clinical risk score with key clinical variables (Fig. 5A). The constructed nomograms were evaluated and the associated calibration curves, ROC curves and DCA curves were plotted. Calibration curves demonstrated good agreement between the nomogram-predicted and observed 1-, 3-, and 5-year overall survival probabilities (Fig. 5B). ROC analysis confirmed the superior predictive accuracy of the nomogram over the risk score alone and other clinical factors, with time-dependent AUCs of 0.759, 0.734, and 0.784 for 1-, 3-, and 5-year survival, respectively (Fig. 5C, E). Decision curve analysis (DCA) further demonstrated that the nomogram provided a greater net benefit across most practical threshold probabilities compared to the risk score and other clinical factors, underscoring its clinical utility (Fig. 5F).

Fig. 5.

Nomogram Development and Model Assessment. A The nomogram integrates risk scores with key clinical determinants of cancer progression, including gender, overall stage, and TNM classification. B Calibration curves demonstrated that the developed nomograms accurately predicted patient survival at 1-, 3-, and 5-year intervals. C-E ROC curve analysis revealed superior predictive accuracy of the nomogram compared to using risk scores or other clinical information alone, achieving AUC values of 0.759, 0.734, and 0.784 for 1-, 3-, and 5-year follow-ups, respectively. F Decision Curve Analysis illustrated that, in terms of practical clinical decision-making, the nomograms provided greater value than reliance on risk scores or other clinical data alone

Risk gene expression and pattern of mutation in hepatocellular carcinoma

To validate the synergistic effects of cuproptosis and macrophage polarization in hepatocellular carcinoma (HCC) progression, we analyzed their interactions using single-cell transcriptomics and spatial transcriptomics technologies, leveraging gene set scores associated with cuproptosis and macrophage polarization. In the spatial transcriptomics of hepatocellular carcinoma, regions with high scores are primarily concentrated in endothelial cells, fibroblasts, and other cell types, and there is significant overlap in the regions where these processes are active (Fig. 6A,B). At the single-cell level, we quantified pathway activities and assessed their interaction using a regression model. Notably, in macrophages, a significant negative interaction was observed (coefficient = -0.071, p = 4.21e-07), suggesting that cuproptosis may antagonize macrophage polarization. In other cell types with significant interaction effects, such as B cells (interaction term = − 0.079, p = 0.005), CD8 + T cells (− 0.040, p = 0.003), fibroblasts (− 0.052, p = 0.003), and monocytes (− 0.076, p < 0.001), copper oxidative stress scores showed a significant negative correlation with polarization scores, suggesting that these cell types may share certain inhibitory regulatory mechanisms (Figure S2, Table S9). We next characterized the six risk genes by analyzing their mutational status, expression characteristics, and spatial distribution in HCC. Analysis of TCGA data revealed that GLIPR2, ANP32E, LIPT1, ARSK, and PGAM1 were upregulated in HCC tumors, while ALAD was downregulated compared to normal tissues (Fig. 6C). Among these, ANP32E, GLIPR2, LIPT1, and ALAD were significantly associated with patient prognosis (Figs. 6D, S3). At the single-cell level, the expression patterns of these genes showed that GLIPR2 was expressed predominantly in T cell, monocyte, macrophage, endothelial cell, and fibroblast subpopulations, ANP32E was highly expressed predominantly in the T cell subpopulation, but also in other subpopulations, LIPT1 was expressed in T cells and macrophages, ALAD was expression, ARSK was more expressed in T cell and fibroblast subpopulations, and PGAM1 was highly expressed in several cell subpopulations (Fig. 6E). And immunohistochemical images sourced from the HPA database showed the expression of risk genes in hepatocellular carcinoma consistent with our analysis (Figures S4, S5). Characterization of the spatial transcriptome data analysis indicated spatial expression profiles of risk genes except for ARSK (Figure S6).

Fig. 6.

Expression and mutation patterns of risk genes at the single-cell and bulk transcriptome levels in HCC. A Spatial visualization of cuproptosis gene scores and macrophage polarization-related gene scores in spatial transcriptomics. B The corresponding annotation distribution results of cells under spatial transcriptomics. C Box plot comparing the expression levels of the 6 risk genes in the normal tissues and tumor tissues. D Kaplan-Meier survival curve showing the overall survival rate difference between the high-expression GLPR2 gene group versus the low expression group in terms of overall survival. P value of 0.020 for the log-rank test indicates a significant difference in survival between the two groups. E UMAP downscaling plots show the distribution of the samples in two dimensions of the expression of the 6 risk genes genes at the single-cell level. The color shades indicate the level of gene expression

Expression of ARSK in hepatocellular carcinoma

To further define the prognostic role of the risk genes, we performed multifactorial Cox regression analysis, which identified ALAD and ARSK as independent prognostic factors for HCC (Fig. 7A). We constructed a protein-protein interaction network for ARSK, revealing key interactions with SULF1, SULF2, ARSA, and GALNS (Fig. 7B). Pan-cancer analysis revealed that ARSK expression was significantly dysregulated in multiple cancer types, including HCC, breast cancer, colorectal cancer, and lung cancer, compared to normal tissues (Fig. 7C). Elevated ARSK expression was associated with poorer overall patient prognosis (Fig. 7D). Furthermore, ARSK expression levels varied significantly according to pathological stage and TP53 mutation status (Fig. 7E, F). Spatial transcriptomic analysis localized ARSK expression primarily to regions enriched with fibroblasts and T cells, a finding consistent with our single-cell RNA sequencing data (Fig. 7G). We validated the expression patterns of all six risk genes (GLIPR2, ANP32E, LIPT1, ALAD, ARSK, and PGAM1) by comparing their levels in HCC cell lines versus a normal hepatocyte cell line (Fig. 7H).

Fig. 7.

Expression of ARSK in hepatocellular carcinoma. A Forest plot showing the Hazard Ratio (Hazard Ratio) and its 95% confidence intervals for six risk genes (GLIPR2, ANP32E, LIPT1, ALAD, ARSK, PGAM1). The number of samples (N) is labeled next to the Hazard Ratio for each gene, as well as the corresponding p-value. B Protein Interaction Network Diagram showing the interaction between ARSK and other proteins (e.g. SUMF1, IDUA, ARSA, ARSB, GALN, ARS, HYAL, etc.) C The plot showing the pattern of ARSK gene expression in different types of tumors, with the color changing from blue (low expression) to red (high expression). D The Kaplan-Meier curve demonstrated a significant difference in survival rates between the high-expression and low-expression groups of ARSK (Log-rank P = 0.0012). E Box and line plot showing the expression level of ARSK gene in hepatocellular carcinoma in different individual cancer stages (Stage1 to Stage4). F Box and line plot showing the expression level of ARSK gene in hepatocellular carcinoma in different TP53 mutation status (wild type and mutant). G The spatial transcriptome data results in the expression of the ARSK gene at spatial locations in different cell types. H qPCR validation results of six risk genes in liver cancer cell lines and normal cell lines. ****for P < 0.0001, ***for P < 0.001, **for P < 0.01, and *for P < 0.05

Discussion

Hepatocellular carcinoma, the incidence of which is increasing globally, has seen great advances in hepatocellular carcinoma immunologic research and related therapies over the past few years [19]. Emerging evidence suggests that copper homeostasis may interact with immune regulation, as copper is an essential cofactor for various enzymatic processes [20], recent studies reveal that excessive copper accumulation can paradoxically improve cancer prognosis through cuproptosis induction, even in the context of immunosuppression. Copper overload triggers cuproptosis via lipoylated protein aggregation, a mechanism leveraged by copper-delivery nanoparticles to eradicate dormant tumors, albeit with CD4 + T cell suppression [21]. Prognostic benefits in select cancers may reflect dominant tumoricidal effects over immunosuppression, especially in mitochondria-dependent malignancies [22]. It is crucial to distinguish cuproptosis from other regulated cell death pathways. Unlike ferroptosis, which is driven by iron-dependent lipid peroxidation or apoptosis, cuproptosis is triggered by copper-induced aggregation of mitochondrial lipoylated proteins without major ROS involvement [5, 23]. A key intersection lies in their shared dependency on mitochondrial metabolism and glutathione (GSH). The ability of drugs like sorafenib to co-induce both ferroptosis and cuproptosis via GSH depletion underscores their mechanistic synergy, offering a potent combined therapeutic avenue in cancer [5, 23–26]. In the tumor microenvironment, macrophage polarization demonstrates dual roles: M1-type macrophages generally exhibit anti-tumor effects, whereas M2-type macrophages often promote cancer progression [27]. Notably, the potential interplay among copper metabolism, macrophage polarization, and HCC progression remains largely unexplored, warranting systematic investigation.

In this study, we explored the correlation between CRMGs and liver cancer. Six risk genes were screened by univariate Cox regression analysis and machine learning algorithms. After constituting our prognostic risk prediction model, we categorized patients into high-risk and low-risk groups, and there were significant prognostic differences between the two groups. In addition, our model successfully predicted the patients’ tumor microenvironment, sensitivity to immunotherapy, and response to chemotherapy. Through GSVA enrichment analysis, we found that there were significant differences in the pathways enriched to biological processes such as fatty acid metabolism, cholesterol homeostasis, and WNT β-catenin signaling pathway between the two groups of patients. Dysregulation of fatty acid metabolism and cholesterol metabolism and are important risk factors for liver cancer development [28, 29]. The Wnt/β-catenin signaling pathway plays an essential role in various physiological functions, including development, tissue homeostasis, and cell proliferation, and its aberrant regulation significantly contributes to hepatocellular carcinoma initiation and progression [30]. During the analysis, we found significant differences in immune cell infiltration between the two groups of patients, especially in the high-risk patients, where M1 TAM was significantly higher than that of patients in the low-risk group, which is consistent with the results of related studies [31, 32]. This observation aligns with emerging insights into the role of copper metabolism in shaping the immune landscape. Copper metabolism is critically involved in regulating macrophage function and polarization within the tumor microenvironment. It can drive macrophages toward a pro-inflammatory M1 phenotype, which exhibits anti-tumor activity, and promote the secretion of factors that enhance T cell infiltration [33–35]. Innovative strategies, such as copper-based nanovaccines, have been designed to leverage this mechanism. These approaches can effectively reprogram immunosuppressive M2-type tumor-associated macrophages into tumor-fighting M1-type macrophages, thereby activating potent anti-tumor immune responses [36].

To refine treatment decisions and improve prognostic accuracy after immunotherapy, it is essential to better understand heterogeneity in patient response. In this study, we identified ten risk genes in hepatocellular carcinoma (HCC) using univariate Cox regression and machine learning methods. These genes help predict patient prognosis and reflect differences in therapeutic benefit across prognostic subgroups. Given that many HCC cases are diagnosed at advanced stages and often develop drug resistance—leading to poor outcomes—there is a clear need for precise, individualized treatment strategies. Kaplan–Meier analysis showed significantly higher survival in the low-risk group compared to the high-risk group, enabling early identification of high-risk patients and supporting risk score-based stratification: high-risk patients may benefit from more aggressive therapy and close monitoring, whereas low-risk patients can follow standard protocols. Such an approach aims to optimize clinical outcomes in HCC. Furthermore, the risk stratification framework revealed a computational association between cupping-associated macrophage features and tumor immune landscape. Based on TIDE algorithm analysis, we observed differences in immunotherapy susceptibility among risk groups. However, while variations in TIDE scores and immune checkpoint expression suggest a link between our risk model and immunosuppression, the clinical predictive utility of TIDE—which estimates immune evasion from transcriptomic features—remains debated in HCC. These findings do not yet constitute clinical evidence of immunotherapy response and require validation in future prospective trials.

The past decade has witnessed an exponential expansion in hepatocellular carcinoma (HCC) immunotherapy research, fundamentally reshaping the treatment paradigm and positioning immune checkpoint inhibitors as a cornerstone for advanced-stage management [37, 38]. Our comprehensive analysis also revealed that 57 of the 79 immune risk genes previously reported exhibited significant differential expression across patients with varying risk levels. Correspondingly, drug sensitivity predictions indicated marked disparities in responses to frontline HCC therapeutics, including sorafenib and cisplatin. Single-cell and spatial transcriptomic analyses indicate that these two biological processes interact, with regression models uncovering significant negative interactions across diverse immune cell populations. These bioinformatic findings suggest potential biological links between copper metabolism dysregulation and immune evasion mechanisms in HCC progression. Our analysis, along with emerging evidence, suggests that cuproptosis plays a significant role in tumor immunity. It can enhance the response to immune checkpoint blockade through two key mechanisms: by modulating the expression of immune checkpoints like PD-L1, and by activating the cGAS-STING pathway to stimulate robust anti-tumor immunity [39–41]. This provides a strong rationale for combining cuproptosis inducers with immunotherapy to improve outcomes in HCC. While the identified immune checkpoint gene expression variations across risk subgroups provide insights into possible differential responses to immunotherapy, it is important to emphasize that these computational predictions require confirmation in clinically annotated immunotherapy cohorts. The integration of macrophage polarization signatures with immune checkpoint profiles offers a multidimensional perspective on HCC heterogeneity, which may inform future investigations into personalized treatment strategies. Our analytical framework demonstrates how systems biology approaches can identify candidate biomarkers worthy of functional validation in copper metabolism-immune interaction studies. By combining the expression profiles of immune checkpoint genes with risk scores, the model has the potential to help clinicians develop more precise and personalized treatment regimens, thereby realizing the potential for prognosis and quality of life for HCC patients.

The six risk genes that constitute the risk stratification framework (GLIPR2, ANP32E, LIPT1, ALAD, PGAM1, and ARSK) promote the progression of hepatocellular carcinoma through the synergistic effects of metabolic reprogramming, anti-apoptotic action, and immune microenvironment regulation. Among these, ARSK stands out as a particularly promising candidate biomarker due to its significant overexpression in HCC tissue and its close association with poor patient prognosis. ARSK encodes a sulfatase that participates in the hydrolysis of sulfates in steroids, carbohydrates, and proteoglycans, playing a key role in hormone regulation and extracellular matrix remodeling [42]. Its overexpression in HCC suggests that it may be involved in tumor-promoting metabolism and immune dysfunction. Notably, our prognostic analysis revealed that high ARSK expression was associated with poorer survival outcomes and was independent of other risk factors in multivariate Cox regression. Interestingly, ARSK expression was also significantly correlated with immune checkpoint molecules, suggesting that it may influence sensitivity to immunotherapy. Single-cell and spatial transcriptomics analyses further localized ARSK expression to T cell and fibroblast subpopulations, indicating its involvement in shaping the tumor immune microenvironment.

The remaining five genes enhance different but complementary tumorigenic mechanisms. PGAM1 and LIPT1 drive metabolic reprogramming, with PGAM1 enhancing the glycolytic pathway and LIPT1 promoting lipid synthesis through the PPARγ/AKT signaling pathway [43, 44]. In contrast, ALAD acts as a protective factor that maintains porphyrin metabolism and metal ion homeostasis, and its deficiency may disrupt copper-dependent cell death pathways [45–47]. Additionally, ANP32E and GLIPR2 are involved in cell death resistance and microenvironment adaptation. ANP32E inhibits ferroptosis through the p53/SLC7A11 axis [48] and promotes chemotherapy resistance in various cancers [49, 50], while GLIPR2 exhibits hypoxia-dependent pro-metastatic activity in HCC [51, 52]. The convergence of these genes suggests a pattern: metabolic dysfunction (PGAM1/LIPT1/ALAD) drives tumor growth while generating vulnerabilities that are counteracted by survival mechanisms mediated by ANP32E and GLIPR2. We further hypothesize that ARSK introduces a critical bridge between this internal stress landscape and external immune evasion. It potentially translates the underlying metabolic perturbations—such as disrupted copper homeostasis and lipoylation pathways—into an immunosuppressive microenvironment, a notion supported by its spatial colocalization with immune cells and its association with immune checkpoint genes. However, the spatiotemporal dynamics of how ARSK modulates immune checkpoint molecular pathways through sulfuric acid metabolism and its synergistic interactions with PGAM1/LIPT1/ALAD within multi-omics networks remain unvalidated. Further analysis using clinical samples and spatial multi-omics technologies is required.

Limitation

This study helps to provide personalized treatment and clinical decision making for patients with HCC. However, it has some limitations. However, there are several limitations that should be recognized. First, complete reliance on public databases needs to be validated by institutional cohorts and prospective clinical studies. Second, although these genes have shown prognostic predictive value, their clinical translation still requires processes to achieve, such as constructing multicenter molecular typing cohorts, advancing pharmacological intervention studies (e.g., development of variant inhibitors/synthesis of lethal strategies) at key nodes (e.g., ANP32E/PGAM1), and realizing closed-loop validation of the biomarkers to precision therapy. Finally, the clinical relevance of our risk score, especially its ability to predict response to immunotherapy, remains to be tested in prospective trials. In the future, we will integrate additional data from our institution and conduct additional experiments to confirm these findings.

Conclusion

We constructed a prognostic model based on cuproptosis-associated macrophage polarization genes capable of stratifying hepatocellular carcinoma patients into risk cohorts with different clinical outcomes and treatment sensitivities. The model demonstrated potential clinical applications by identifying high-risk individuals who may benefit from intensive immunotherapy or chemotherapy regimens, enabling personalized treatment strategies and optimized monitoring protocols. Importantly, our analysis revealed six candidate biomarkers whose expression patterns mechanistically link copper-dependent cell death to macrophage plasticity in the tumor microenvironment. If validated, these biomarkers would facilitate early detection of disease progression and guide precision treatment decisions. While experimental validation of the correlation between gene function and drug response will be validated in the future, the robustness of the model is supported by multi-cohort validation and consistency with public datasets. Further research will focus on elucidating the mechanistic role of CRMGs and translating this risk stratification framework into clinical practice.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Material preparation and data collection were performed by X.D., Y.L., B.L. and J.C. Data analysis was completed by X.D. The first draft of the manuscript was written by X.D., Y.L. and all authors commented on previous versions of the manuscript. Review of articles and revisions by X.D. and L.T. All authors read and approved the final manuscript.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Data availability

The data for this study were obtained from the TCGA, GEO and ICGC databases, all of which are publicly available. The hepatocellular carcinoma single cell dataset is GSE149614 from the GEO database. ICGC LIRI-JP transcriptomic expression data were downloaded as a validation cohort. Publicly available spatial transcriptome data are available in the Genome Sequence Archive (GSA) under accession number HRA000437, and more details about ST data are available at https://ngdc.cncb.ac.cn/gsa-human/browse/HRA000437.

Declarations

Ethics approval and consent to participate

The data used in this study were obtained from publicly available data in TCGA, ICGC, GSA and GEO databases and did not involve ethical approval or consent to participate.

Consent for publication

All authors have consented to the publication of the manuscript.

Competing interests

The authors declare no competing interests.